Chinese artificial intelligence firm DeepSeek rocked markets this week with claims its new AI model outperforms OpenAI’s and cost a fraction of the price to build.
The assertions — specifically that DeepSeek’s large language model cost just $5.6 million to train — have sparked concerns over the eyewatering sums that tech giants are currently spending on computing infrastructure required to train and run advanced AI workloads.
But not everyone is convinced by DeepSeek’s claims.
CNBC asked industry experts for their views on DeepSeek, and how it actually compares to OpenAI, creator of viral chatbot ChatGPT which sparked the AI revolution.
What is DeepSeek?
Last week, DeepSeek released R1, its new reasoning model that rivals OpenAI’s o1. A reasoning model is a large language model that breaks prompts down into smaller pieces and considers multiple approaches before generating a response. It is designed to process complex problems in a similar way to humans.
DeepSeek was founded in 2023 by Liang Wenfeng, co-founder of AI-focused quantitative hedge fund High-Flyer, to focus on large language models and reaching artificial general intelligence, or AGI.
AGI as a concept loosely refers to the idea of an AI that equals or surpasses human intellect on a wide range of tasks.
Much of the technology behind R1 isn’t new. What is notable, however, is that DeepSeek is the first to deploy it in a high-performing AI model with — according to the company — considerable reductions in power requirements.
“The takeaway is that there are many possibilities to develop this industry. The high-end chip/capital intensive way is one technological approach,” said Xiaomeng Lu, director of Eurasia Group’s geo-technology practice.
“But DeepSeek proves we are still in the nascent stage of AI development and the path established by OpenAI may not be the only route to highly capable AI.”
How is it different from OpenAI?
DeepSeek has two main systems that have garnered buzz from the AI community: V3, the large language model that unpins its products, and R1, its reasoning model.
Both models are open-source, meaning their underlying code is free and publicly available for other developers to customize and redistribute.
DeepSeek’s models are much smaller than many other large language models. V3 has a total of 671 billion parameters, or variables that the model learns during training. And while OpenAI doesn’t disclose parameters, experts estimate its latest model to have at least a trillion.
In terms of performance, DeepSeek says its R1 model achieves performance comparable to OpenAI’s o1 on reasoning tasks, citing benchmarks including AIME 2024, Codeforces, GPQA Diamond, MATH-500, MMLU and SWE-bench Verified.
Read more DeepSeek coverage
In a technical report, the company said its V3 model had a training cost of only $5.6 million — a fraction of the billions of dollars that notable Western AI labs such as OpenAI and Anthropic have spent to train and run their foundational AI models. It isn’t yet clear how much DeepSeek costs to run, however.
If the training costs are accurate, though, it means the model was developed at a fraction of the cost of rival models by OpenAI, Anthropic, Google and others.
Daniel Newman, CEO of tech insight firm The Futurum Group, said these developments suggest “a massive breakthrough,” although he shed some doubt on the exact figures.
“I believe the breakthroughs of DeepSeek indicate a meaningful inflection for scaling laws and are a real necessity,” he said. “Having said that, there are still a lot of questions and uncertainties around the full picture of costs as it pertains to the development of DeepSeek.”
Meanwhile, Paul Triolio, senior VP for China and technology policy lead at advisory firm DGA Group, noted it was difficult to draw a direct comparison between DeepSeek’s model cost and that of major U.S. developers.
“The 5.6 million figure for DeepSeek V3 was just for one training run, and the company stressed that this did not represent the overall cost of R&D to develop the model,” he said. “The overall cost then was likely significantly higher, but still lower than the amount spent by major US AI companies.”
DeepSeek wasn’t immediately available for comment when contacted by CNBC.
Comparing DeepSeek, OpenAI on price
DeepSeek and OpenAI both disclose pricing for their models’ computations on their websites.
DeepSeek says R1 costs 55 cents per 1 million tokens of inputs — “tokens” referring to each individual unit of text processed by the model — and $2.19 per 1 million tokens of output.
In comparison, OpenAI’s pricing page for o1 shows the firm charges $15 per 1 million input tokens and $60 per 1 million output tokens. For GPT-4o mini, OpenAI’s smaller, low-cost language model, the firm charges 15 cents per 1 million input tokens.
Skepticism over chips
DeepSeek’s reveal of R1 has already led to heated public debate over the veracity of its claim — not least because its models were built despite export controls from the U.S. restricting the use of advanced AI chips to China.
DeepSeek claims it had its breakthrough using mature Nvidia clips, including H800 and A100 chips, which are less advanced than the chipmaker’s cutting-edge H100s, which can’t be exported to China.
However, in comments to CNBC last week, Scale AI CEO Alexandr Wang, said he believed DeepSeek used the banned chips — a claim that DeepSeek denies.
Nvidia has since come out and said that the GPUs that DeepSeek used were fully export-compliant.
The real deal or not?
Industry experts seem to broadly agree that what DeepSeek has achieved is impressive, although some have urged skepticism over some of the Chinese company’s claims.
“DeepSeek is legitimately impressive, but the level of hysteria is an indictment of so many,” U.S. entrepreneur Palmer Luckey, who founded Oculus and Anduril wrote on X.
“The $5M number is bogus. It is pushed by a Chinese hedge fund to slow investment in American AI startups, service their own shorts against American titans like Nvidia, and hide sanction evasion.”
Seena Rejal, chief commercial officer of NetMind, a London-headquartered startup that offers access to DeepSeek’s AI models via a distributed GPU network, said he saw no reason not to believe DeepSeek.
“Even if it’s off by a certain factor, it still is coming in as greatly efficient,” Rejal told CNBC in a phone interview earlier this week. “The logic of what they’ve explained is very sensible.”
However, some have claimed DeepSeek’s technology might not have been built from scratch.
“DeepSeek makes the same mistakes O1 makes, a strong indication the technology was ripped off,” billionaire investor Vinod Khosla said on X, without giving more details.
It’s a claim that OpenAI itself has alluded to, telling CNBC in a statement Wednesday that it is reviewing reports DeepSeek may have “inappropriately” used output data from its models to develop their AI model, a method referred to as “distillation.”
“We take aggressive, proactive countermeasures to protect our technology and will continue working closely with the U.S. government to protect the most capable models being built here,” an OpenAI spokesperson told CNBC.
Commoditization of AI
However the scrutiny surrounding DeepSeek shakes out, AI scientists broadly agree it marks a positive step for the industry.
Yann LeCun, chief AI scientist at Meta, said that DeepSeek’s success represented a victory for open-source AI models, not necessarily a win for China over the U.S. Meta is behind a popular open-source AI model called Llama.
“To people who see the performance of DeepSeek and think: ‘China is surpassing the US in AI.’ You are reading this wrong. The correct reading is: ‘Open source models are surpassing proprietary ones’,” he said in a post on LinkedIn.
“DeepSeek has profited from open research and open source (e.g. PyTorch and Llama from Meta). They came up with new ideas and built them on top of other people’s work. Because their work is published and open source, everyone can profit from it. That is the power of open research and open source.”
Synopsys logo is seen displayed on a smartphone with the flag of China in the background.
Sopa Images | Lightrocket | Getty Images
The U.S. government has rescinded its export restrictions on chip design software to China, U.S.-based Synopsys announced Thursday.
“Synopsys is working to restore access to the recently restricted products in China,” it said in a statement.
The U.S. had reportedly told several chip design software companies, including Synopsys, in May that they were required to obtain licenses before exporting goods, such as software and chemicals for semiconductors, to China.
The U.S. Commerce Department did not immediately respond to a request for comment from CNBC.
The news comes after China signaled last week that they are making progress on a trade truce with the U.S. and confirmed conditional agreements to resume some exchanges of rare earths and advanced technology.
The Datadog stand is being displayed on day one of the AWS Summit Seoul 2024 at the COEX Convention and Exhibition Center in Seoul, South Korea, on May 16, 2024.
Chris Jung | Nurphoto | Getty Images
Datadog shares were up 10% in extended trading on Wednesday after S&P Global said the monitoring software provider will replace Juniper Networks in the S&P 500 U.S. stock index.
S&P Global is making the change effective before the beginning of trading on July 9, according to a statement.
Computer server maker Hewlett Packard Enterprise, also a constituent of the index, said earlier on Wednesday that it had completed its acquisition of Juniper, which makes data center networking hardware. HPE disclosed in a filing that it paid $13.4 billion to Juniper shareholders.
Over the weekend, the two companies reached a settlement with the U.S. Justice Department, which had sued in opposition to the deal. As part of the settlement, HPE agreed to divest its global Instant On campus and branch business.
While tech already makes up an outsized portion of the S&P 500, the index has has been continuously lifting its exposure as the industry expands into more areas of society.
Stocks often rally when they’re added to a major index, as fund managers need to rebalance their portfolios to reflect the changes.
New York-based Datadog went public in 2019. The company generated $24.6 million in net income on $761.6 million in revenue in the first quarter of 2025, according to a statement. Competitors include Cisco, which bought Splunk last year, as well as Elastic and cloud infrastructure providers such as Amazon and Microsoft.
Datadog has underperformed the broader tech sector so far this year. The stock was down 5.5% as of Wednesday’s close, while the Nasdaq was up 5.6%. Still, with a market cap of $46.6 billion, Datadog’s valuation is significantly higher than the median for that index.
A representation of cryptocurrency Ethereum is placed on a PC motherboard in this illustration taken on June 16, 2023.
Dado Ruvic | Reuters
Stocks tied to the price of ether, better known as ETH, were higher on Wednesday, reflecting renewed enthusiasm for the crypto asset amid a surge of interest in stablecoins and tokenization.
“We’re finally at the point where real use cases are emerging, and stablecoins have been the first version of that at scale but they’re going to open the door to a much bigger story around tokenizing other assets and using digital assets in new ways,” Devin Ryan, head of financial technology research at Citizens.
On Tuesday, as bitcoin ETFs snapped a 15-day streak of inflows, ether ETFs saw $40 million in inflows led by BlackRock’s iShares Ethereum Trust. ETH ETFs came back to life in June after much concern that they were becoming zombie funds.
The price of the coin itself was last higher by 5%, according to Coin Metrics, though it’s still down 24% this year.
Ethereum has been struggling with an identity crisis fueled by uncertainty about the network’s value proposition, weaker revenue since its last big technical upgrade and increasing competition from Solana. Market volatility, driven by geopolitical uncertainty this year, has not helped.
The Ethereum network’s smart contracts capability makes it a prominent platform for the tokenization of traditional assets, which includes U.S. dollar-pegged stablecoins. Fundstrat’s Tom Lee this week called Ethereum “the backbone and architecture” of stablecoins. Both Tether (USDT) and Circle‘s USD Coin (USDC) are issued on the network.
BlackRock’s tokenized money market fund (known as BUIDL, which stands for USD Institutional Digital Liquidity Fund) also launched on Ethereum last year before expanding to other blockchain networks.
Tokenization is the process of issuing digital representations on a blockchain network of publicly traded securities, real world assets or any other form of value. Holders of tokenized assets don’t have outright ownership of the assets themselves.
The latest wave of interest in ETH-related assets follows an announcement by Robinhood this week that it will enable trading of tokenized U.S. stocks and ETFs across Europe, after a groundswell of interest in stablecoins throughout June following Circle’s IPO and the Senate passage of its proposed stablecoin bill, the GENIUS Act.
Ether, which turns 10 years old at the end of July, is sitting about 75% off its all-time high.
Don’t miss these cryptocurrency insights from CNBC Pro: